Advances in Engineering Software 77 (2014) 35–47

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

An effective hybrid teaching–learning-based optimization algorithm

for permutation flow shop scheduling problem
Zhanpeng Xie, Chaoyong Zhang ⇑, Xinyu Shao, Wenwen Lin, Haiping Zhu
State Key Laboratory of Digital Manufacturing Equipment & Technology, Huazhong University of Science & Technology (HUST), Wuhan, Hubei 430074, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Permutation flow shop scheduling (PFSP) is among the most studied scheduling settings. In this paper, a
Received 19 March 2014 hybrid Teaching–Learning-Based Optimization algorithm (HTLBO), which combines a novel teaching–
Received in revised form 27 June 2014 learning-based optimization algorithm for solution evolution and a variable neighborhood search
Accepted 13 July 2014
(VNS) for fast solution improvement, is proposed for PFSP to determine the job sequence with minimiza-
Available online 24 August 2014
tion of makespan criterion and minimization of maximum lateness criterion, respectively. To convert the
individual to the job permutation, a largest order value (LOV) rule is utilized. Furthermore, a simulated
Keywords:
annealing (SA) is adopted as the local search method of VNS after the shaking procedure. Experimental
Permutation flow shop scheduling problem
Teaching-learning-based optimization
comparisons over public PFSP test instances with other competitive algorithms show the effectiveness
algorithm of the proposed algorithm. For the DMU problems, 19 new upper bounds are obtained for the instances
Variable neighborhood search with makespan criterion and 88 new upper bounds are obtained for the instances with maximum late-
Simulated annealing ness criterion.
Makespan Ó 2014 Elsevier Ltd. All rights reserved.
Maximum lateness

1. Introduction and Cj its completion time at the last machine of the shop. Lenstra
et al. [7] proved that the two-machine flow shop with maximum
In permutation flow shop scheduling problems, n jobs N = N1, lateness is NP-complete.
N2, . . ., Nn have to be processed on a set of m machines M = M1, Approaches for PFSP can be divided into three categories: exact
M2, . . ., Mm sequentially. Therefore, each job consists of a set of m algorithms, heuristics and meta-heuristics. Exact algorithms, such
operations Oj = {Oj1, . . ., Ojm}. Each operation has a given processing as branch-and-bound method, dynamic programming and mathe-
time denoted by Pi,j (i = 1, 2, . . ., m, j = 1, 2, . . ., n). At any time, each matical programming, have been successfully applied in solving
machine can process at most one job and each job can be processed small instances [8–10]. However, they could not obtain promising
by at most one machine. Once the processing of a job on a machine results in a reasonable time for medium or large instances. As for
has started, it must be completed without interruption. The the heuristics, a feasible solution is generally built based on some
sequence in which the jobs to be processed are identical for each constructive operations with a fast process, while the solution is
machine. Thus there is n! possible processing sequences for the quite not satisfactory [11]. More recently, the meta-heuristic algo-
problem. The minimum completion time, which is known as make- rithms, such as genetic algorithm (GA), simulated annealing (SA),
span or Cmax, is the most commonly studied objective of PFSP [1]. tabu search (TS), have been given special emphasis for they could
Recently, PFSP with other objectives such as those involving due provide high-quality solutions with reasonable computing times.
dates have drawn significant attention [2–4]. Demirkol et al. [5] In recent decade, an increasing number of research papers focusing
presented extensive sets of randomly generated test problems for on meta-heuristics for PFSP have been published.
the problems of minimizing makespan (Cmax) and maximum late- Teaching–Learning-Based Optimization (TLBO) proposed by Rao
ness (Lmax) in flow shops, generally referred to DMU problems. et al. [12] is a novel efficient optimization method. The main idea
PFSP with the makespan criterion can be denoted as Fm|prmu|Cmax behind TLBO is the simulation of a classical school learning pro-
and has been proved NP-complete [6]. PFSP with the criterion of cess. The advantages of TLBO algorithm such as ease of implemen-
maximum lateness can be denoted as Fm|prmu|Lmax, where tation, immediately accessible for practical applications, speed to
Lj = max{Cj  dj, 0} is the lateness of job j, being dj its due date get the solutions and robustness are shown in the literature
[12,13]. TLBO seems to be a rising star from amongst a number
⇑ Corresponding author. Tel.: +86 13429882087. of meta-heuristics with relatively competitive performances.
E-mail address: [email protected] (C. Zhang). Empirical tests show that TLBO could outperforms the other

http://dx.doi.org/10.1016/j.advengsoft.2014.07.006
0965-9978/Ó 2014 Elsevier Ltd. All rights reserved.
36 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

well-known meta-heuristics regarding constrained benchmark For the maximum lateness criterion, within our knowledge,
functions, constrained mechanical design, and continuous non-lin- only a few of researchers adopted this criterion as the performance
ear numerical optimization problems [13]. However, applications measure of proposed algorithms. Some researchers studied on the
of TLBO for discrete combinatorial optimization problems are still two-machine flow shop scheduling with maximum lateness crite-
limited. rion [47]. Tasgetiren et al. [38] first introduced a particle swarm
In this paper, a novel hybrid Teaching–Learning-Based Optimi- optimization for maximum lateness minimization in permutation
zation algorithm (HTLBO) is proposed for PFSP to optimize two flow shop scheduling problem based on the DMU benchmark prob-
objectives: the makespan and maximum lateness of jobs. The lems. Since then, some novel meta-heuristics have been proposed
paper is organized as follows: Section 2 presents the literature to deal with the objective of maximum lateness. Zheng and Yamas-
review about PFSP problem; Section 3 provides the description of hiro [11] designed a new quantum differential evolutionary algo-
the PFSP; Section 4 describes implementation details of the HTLBO rithm, this algorithm based on the basic quantum-inspired
for PFSP; Section 5 shows the computational results and compari- evolutionary algorithm (QEA). Li and Yin [48] suggested an opposi-
sons with other competitive algorithms; and Section 6 concludes tion-based differential evolution algorithm to solve PFSP with the
the paper. criteria of makespan and maximum lateness.

3. Description of the permutation flow shop problem

2. Literature review
Schedule p is a permutation of the n jobs, which can be denoted
The makespan criterion for flow shop scheduling is still a hot as {p1, p2, . . ., pn}, in which pi e X is the ith (i = 1, 2, . . ., n) job in p. P
topic of research as shown in the recent review by Gupta and is the set of all the permutations of the n jobs. Let ppi ;j represents
Stafford Jr. [14]. Tseng et al. [15] introduced that the exact algo- the processing time of job pi on machine j and C(pi, m) represents
rithms are sensible to the number of machines. As a result, a the completion time of job pi on machine m. Then the completion
wealth of literature on heuristic and meta-heuristic methods for time for the n-job, m-machine problem can be calculated as
the PFSP problem and makespan criterion was published. Heuris- follows:
tics for the makespan minimization problem have been proposed
by Palmer [16], Campbell et al. [17], Dannenbring [18], Nawaz Cðp1 ; 1Þ ¼ pp1 ; 1 ð1Þ
et al. [19], Taillard [20], Framinan and Leisten [21] and Framinan
et al. [22], Li et al. [23], Laha and Chakraborty [24]. Among these Cðpi ; 1Þ ¼ Cðpi1 ; 1Þ þ ppi ;1 ; i ¼ 2; . . . ; n ð2Þ
existing heuristics, the Nawaz–Enscore–Ham (NEH) heuristic has
been proved one of the most successful constructive heuristics Cðp1 ; jÞ ¼ Cðp1 ; j  1Þ þ pp1 ;j ; j ¼ 2; . . . ; m ð3Þ
which can obtain comparable results against some modern con-
structive methods according to the results of the computational Cðpi ; jÞ ¼ maxðCðpi1 ; jÞ; Cðpi ; j  1ÞÞ þ ppi ;j ; i ¼ 2; . . . ; n; j ¼ 2; . . . ; m
evaluation given by Ruiz and Maroto [25]. The meta-heuristics ð4Þ
include simulated annealing [26,27], tabu search [28–31], genetic
algorithms [32,33], ant colony optimization [34], iterated local The makespan of a permutation p can be formally defined as the
search [35], iterated greedy methods [36,37], particle swarm opti- completion time pn of the last job on the last machine m, so the
mization [38–40], differential evolution algorithm [41] and so on. makespan is defined as:
Based on these meta-heuristics, some hybrid algorithms were pro- C max ðpÞ ¼ Cðpn ; mÞ ð5Þ
posed, which had been demonstrated effective according to the
computational results on some well-known benchmark problems. The PFSP with the makespan criterion is to find the optimal per-
Chang et al. [42] proposed a hybrid genetic-immune algorithm, in mutation p* in the set of all permutation:
which the regular genetic algorithm is applied in the first stage to C max ðp Þ  Cðpn ; mÞ; 8p 2 P ð6Þ
rapidly evolve and when the processes are converged up to a pre-
defined iteration then artificial immune system is introduced to As for the flow shop scheduling with the due date constraint, let
hybridize GA in the second stage. Among hybrid algorithms, the L(pi) denoted the lateness of jobs pi and can be defined as:
most popular strategy is to hybridize a meta-heuristic with a local Lðpi Þ ¼ Cðpi ; mÞ  dðpi Þ ð7Þ
search method. Murata et al. [43] showed two hybrid algorithms:
genetic local search and genetic simulated annealing. Nearchou Maximum lateness Lmax(p) of a permutation can be defined as:
[44] designed a hybrid simulated annealing algorithm for solving Lmax ðpÞ ¼ maxðCðpi ; mÞ  dðpi ÞÞ ð8Þ
the flow shop scheduling problem. In this algorithm, an iterated
*
hill climbing procedure is applied on the population of schedules where d(pi) is the due date of jobs pi. The optimal solution p should
during the annealing process. Pan et al. [41] introduced a new satisfy the following criterion:
and novel referenced local search procedure hybridized with both Lmax ðp Þ  Lmax ðpÞ; 8p 2 P ð9Þ
discrete differential evolution algorithm and iterated greedy algo-
rithm to further improve the solution quality. Ahmadizar [45]
developed a new ant colony optimization algorithm for makespan 4. Hybrid teaching–learning-based optimization algorithm for
minimization in permutation flow shops. In this algorithm, novel PFSP
mechanism is employed in initialization the pheromone trails
based on an initial sequence. Tzeng et al. [46] proposed a hybrid 4.1. Brief introduction of TLBO
estimation of distribution algorithm (EDA) with ant colony system
(ACS). Their algorithm, in each iteration, applies a new filter strat- Teaching–Learning-Based Optimization (TLBO) is a population-
egy and a local search method to update the local best solution based method inspired by the effect of the influence of a teacher on
and, based on the local best solution, generates pheromone trails the output of learners in a class and has been applied to cluster
using a new pheromone-generating rule and applies a solution data [49], design of planar steel frames [50], optimization of two
construction method of ACS to generate members for the next stage thermoelectric cooler [51], job shop scheduling [52] and so
iteration. on. Like other nature-inspired algorithms, TLBO uses a population
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 37

of solutions to proceed to the global solution. In TLBO, the best then /i;3 ¼ 4 and pi;/i;3 ¼ pi;4 ¼ 3; and so on. Thus, we obtain the
individual is regarded as the teacher, and the rest individuals are job permutation pi = [2, 4, 1, 3, 5]. As we can see, LOV rule provides
regarded as students. The process of TLBO is divided into two a simple conversion to makes TLBO applicable to solve PFSP.
parts: teacher phase and learner phase.
During the teacher phase, a teacher wants to bring his or her 4.3. Initial population
learners up to his or her level in terms of knowledge, but the level
the class can reach depends on the capability of the class. In other The initial population is generated randomly and uniformly. A
words, the improvement of students is influenced by the difference vector Xi = {xi1, xi2, . . ., xin} is randomly produced according to the
between the teacher’s knowledge and the qualities of all students. following formula:
Thus the individuals are modified with the following expression:
xij ¼ xmin þ ðxmax  xmin Þ  rand; i ¼ 1; 2; . . . ; NP; j ¼ 1; 2; . . . ; n
X new;i ¼ X old;i þ r  ðX teacher  ðT F  X mean ÞÞ ð10Þ ð14Þ
where Xteacher is the best individual in the population, TF is a teach-
where xmin = 1.0, xmax = 1.0 and rand is an uniform random num-
ing factor, which can be either 1 or 2, Xmean is the current mean
ber between 0 and 1.
value of the individuals, r is a uniform random number between 0
and 1.
4.4. Crossover operator
During the learner phase, learners increase their knowledge by
interaction between themselves. A random individual Xii is
Crossover is a typical operator in genetic algorithm. To enforce
selected for an individual Xi to learn from. Two situations are con-
the performance of the global search, we apply crossover operator
sidered. If Xii is better than Xi, the individual is modified with the
after the general TLBO tenure. To create new individuals, two types
following expression:
of crossover operators are selected: TP (two-point order crossover)
X new;i ¼ X i þ r  ðX ii  X i Þ ð11Þ operator by Murata et al. [43] and PMX (partially mapped cross-
over) operator by Glodberg [55]. Readers can refer to the literature
Otherwise, the individual is modified with the following
we have mentioned for the details of the crossover operators. Gen-
expression:
erally, a new pair will be generated after crossover between an
X new;i ¼ X i þ r  ðX i  X ii Þ ð12Þ existing pair. However, in our HTLBO, crossover is taken between
an individual and his/her previous individual in the population to
Procedures of TLBO are described as follows and shown in
generate a new pair of individuals. Especially, the first individual
Fig. 1:
is manipulated with the last individual. Each individual create a
new pair of individuals. If the better individual between the new
Step 1: Initialization.
pair is also better than the original individual, the original individ-
Step 2: If the termination criterion is not met, repeat the follow-
ual will be replaced by the better individual.
ing steps.
The procedure of crossover operator can be summarized as
Step 3: Teacher phase.
follows:
Step 4: Learner phase.
Step 1: Apply the LOV rule to convert the individual Xi to the job
4.2. Solution representation permutation pi.
Step 2: Choose one type of crossover operators randomly.
TLBO was initially designed to solve continuous optimization Step 3: Generate a pair of new individuals by crossover and find
problems. Hence, the standard TLBO could not be used to solve the better individual between the new pair.
PFSP directly. In order to apply general TLBO to PFSP, the key issue Step 4: If the better individual is better than the original indi-
is to find a suitable mapping between the job permutation and the vidual, then the original individual is replaced by the better
vector of individuals. Bean et al. presented a robust representation individual.
called random keys [53]. Then inspired by this method, a smallest
position value (SPV) rule based on random keys is proposed by Tas- 4.5. Local search using VNS
getiren et al. [38] to convert individual Xi = [xi,1, xi,2, . . ., xi,n] to the
job permutation pi = [pi,1, pi,2, . . ., pi,n]. Qian et al. [54] introduced Variable neighborhood search is an efficient method for solving
a largest order value (LOV) rule based on random key representa- combinatorial and global optimization problems whose basic idea
tion. Li and Yin [48] proposed a largest position value (LPV) rule. is a systematic change of neighborhood both within a descent
In this work, the LOV rule is adopted for TLBO. A simple example phase to find a local optimum and in a perturbation phase to get
is presented in Table 1 to illustrate the LOV rule. out of the corresponding valley [56].
According to LOV rule, all elements of Xi = [xi,1, xi,2, . . ., xi,n] are In a VNS algorithm, a set of neighborhood structures Nk(x), k = 1,
firstly ranked by descending order to get a sequence /i = [/i,1, /i,2, 2, . . ., kmax in which Nk(x) is the kth neighborhood, is first defined.
. . ., /i,n]. Then the job permutation pi is calculated by the following Then, an initial solution x is found and a stopping criterion is deter-
formula: mined. Given the initial solution x, a random point x0 in Nk(x) is
pi;/i;l ¼ l ð13Þ generated. Starting from x0 , a local search is then performed to gen-
erate x00 . If x00 is better than the incumbent best solution x, then
where the dimension l varies from 1 to n. In Table 1, the LOV is illus- x ¼ x00 , and the search returns to N1(x). Otherwise, the search
trated with a simple instance (n = 5), where individual is Xi = [0.05, explores the next neighborhood N2(x). This is repeated until
0.35, 0.67, 0.21, 0.72]. Because xi,2 is the largest value of Xi. So xi,2 k = kmax.
is selected first and assigned the rank value 1. Then xi,4 is selected The procedure of the basic VNS is described as follows:
secondly and assigned rank value 2. In the same way, xi,1, xi,3, xi,5
are assigned the value 3, 4 and 5, respectively. Thus, the sequence (1) Generate an initial solution.
is /i = [3, 1, 4, 2, 5]. According to (13), if l = 1, then /i,1 = 3 and (2) Select the set of neighborhood structures Nk(x), k = 1, 2, . . .,
pi;/i;1 ¼ pi;3 ¼ 1; if l = 2, then /i;2 ¼ 1 and pi;/i;2 ¼ pi;1 ¼ 2; if l = 3, kmax, that will be used in the search.
38 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

Fig. 1. Flowchart of TLBO algorithm.

Table 1 (b) Local search: Apply some local search method with
Solution representation of individual X ti . x0 as initial solution; denote with x00 the so
Job, dimension 1 2 3 4 5 obtained local optimum.
Position, xtij 0.05 0.35 -0.67 0.21 0.72
(c) Move or not: If this local optimum is better than
the incumbent, move there ðx x00 Þ, and continue
Sequence, /tij 3 1 4 2 5
the search with N1(x); otherwise, set k k + 1.
Job, ptij 2 4 1 3 5

The VNS procedure can be coupled with other procedures in

many different ways to improve the performance. In this paper,
(3) Repeat the following until the stopping criterion is met: we adopted a SA as the local search method after shaking proce-
(1) Set k 1; dure to enhance the search capability of VNS. The performance of
(2) Until k = kmax, repeat the following steps: SA depends on several control parameters: the initial temperature
(a) Shaking: Generate a point x0 at random form the T0, the cooling rate Cr and the final temperature Tf [57]. A standard
kth neighborhood of xðx0 2 N k ðxÞÞ. SA procedure offers the possibility of accepting worse neighbor
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 39

solutions in a controlled manner in order to escape from local min- Five kinds of neighborhood structures are selected. These oper-
ima. However, worse neighbor solutions are not accepted in our ators are written as Swap, Forward-insert, Backward-insert, Inverse
algorithm because we focus on the local search capability of SA. and Adjacent-swap and shown in Fig. 3. The details of these neigh-
The pseudo-code of VNS in this paper is shown in Fig. 2. borhood structures are as follows:
Swap: Choose two different positions from a job permutation
randomly and swap them.
Forward-insert: Choose two different positions from a job per-
mutation randomly and insert the back one before the front.
Backward-insert: Choose two different positions from a job per-
mutation randomly and insert the front one before the back.
Inverse: Inverse the subsequence between two different random
positions of a job permutation.
Adjacent-swap: Choose one position from a job permutation
randomly and swap it with the next position of the job permuta-
tion. Especially, if the chosen position is the last position of the
job permutation, swap it with the first position of the job
permutation.

4.6. Re-initialization mechanism

When the algorithm is trapped in local optimum, it is time-con-

suming to step out of the region. To escape from this situation, we
design a re-initialization mechanism. If the best individual of the
population is not improved after searching successive 50 genera-
tions, we reinitialize the population. The re-initialization strategy
Fig. 2. The pseudo-code of VNS. is designed as follows: half of the individuals are replaced by the

Fig. 3. The neighborhood structures for VNS.

40 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

best individual generated in the previous process, and another half generated randomly during the algorithm and can be either 1 or 2,
are constructed randomly. in which 1 corresponds to no increase in the knowledge level and
2 corresponds to complete transfer of knowledge.
4.7. HTLBO description HTLBO is implemented in C++ and tested on a PC with Intel
Core2 Duo 2.0 GHz CPU & 2 GB memory. Car1, car2 through to
The hybrid algorithm must attain a balance between explora- car8 benchmark problems designed by Carlier [58] and Rec01,
tion and exploitation. In HTLBO, the main role of TLBO is to explore Rec03 through to Rec41 designed by Reeves and Yamada [33] are
the searching space, and the general TLBO is modified by introduc- considered. Average percentage relative difference (ARPD) is
ing the LOV rule to convert the individual to the job permutation. adopted to evaluate the performance.
The main role of VNS is to exploit the individual obtained by the
XR  
global TLBO, and five kinds of neighborhood structures are pre- ðSi  Sbest Þ  100
APRD ¼ R ð15Þ
sented to obtain promising results. The re-initialization mecha- Sbest
i¼1
nism is applied to increase the probability of escaping from the
local optimum. In Eq. (15), for each instance, Si denotes the solution generated
Notations are given before the proposed HTLBO description: by a given algorithm, R is the number of replications and Sbest rep-
resents the best known solution for the instance. The lower APRD
NP population size is, the higher the performance is. In the experiment, R = 10 and
Maxiter maximum number of iteration the termination criterion is set as 0.3  m  n seconds maximum
maxT maximum CPU time computation time. The experiment results are analyzed by the
CT limitation of consecutive generations for multi-factor analysis of variance (ANOVA) method. In the experi-
solution not improved ment, the three main hypotheses (normality, homoscedasticity
count record of consecutive generations for solution and independence of the residuals) are checked and accepted.
not improved The p-values in the experiment are all close to zero, so analyzing
Fi the fitness of the ith individual the p-values is useless. Instead, we focus on the F-ratio, which is
bestvalue[i] the objective function value of the ith generation the ratio between variance explained by a factor and the unex-
plained variance. The greater the F-ratio is, the more effect the fac-
tor has on the response variable. Note that the interactions among
more than two factors are not considered, since their F-ratio are
The pseudo-code of the HTLBO algorithm is shown in Fig. 4. quite small. The factor with the greatest F-ratio is first analyzed,
followed by the second one, and so on.
The greatest F-ratio corresponds to the factor T0, and the means
5. Computational results and comparisons
plot with the Least Significant Difference (LSD) interval (at the 95%
confidence level) is given in Fig. 5. Fig. 5 illustrates that HTLBO
In this section, parameters of HTLBO are determined by exper-
with T0 = 10,000 obtains the significantly best performance, while
iments. Performance of HTLBO is evaluated by comparing it with
that with T0 = 100 yields the worst effectiveness.
other existing good algorithms for the considered problem.
The factor Cr has the second greatest F-ratio, and the means plot
with LSD intervals (at the 95% confidence level) is given in Fig. 6. As
5.1. Parameter determination
we can see from Fig. 6, HTLBO with Cr = 0.85 obtains the signifi-
cantly best performance. HTLBO with a small Cr value generates
The parameters of the HTLBO include the size of population NP,
the worst effectiveness, and the reason lies in that few neighbors
the teaching factor TF, the initial temperature T0, the cooling rate Cr
are searched.
and the final temperature Tf. The effect of population size is consid-
The third greatest F-ratio corresponds to the factor NP. The
erable. The large value of NP would improve the solution quality
means plot with LSD intervals (at the 95% confidence level) is given
but increase the CPU time that is not desirable. Initial temperature,
in Fig. 7. In Fig. 7, it can be seen that HTLBO with NP = 40 obtains
final temperatures and cooling rate play important roles in SA algo-
the significantly best performance. NP = 20 yields the worst effec-
rithm. High initial temperature can extend search scope in initial
tiveness and the reason is that diversity of population is insuffi-
steps. However, when the final temperature is fixed, the increase
cient. If we look closely at the plotted means, we find that the
of initial temperature will increase the running time of the algo-
differences in the response variable different NP are low, which
rithm. Low final temperature causes the algorithm to make narrow
supports that our proposed algorithm is robust.
its search scope and find an acceptable solution. As for the cooling
The last factor is TF, and the means plot with LSD intervals (at
scheme, the most commonly used temperature reducing function
the 95% confidence level) is given in Fig. 8. As we can see from
is i.e. Ti = CrTi1 in which Cr is a constant. We also adopt this reduc-
Fig. 8, there is a clear statistically significant difference between
ing function. Typically, Cr is set between 0.75 and 0.95. The
TF = 1 and TF = 2 schemes and the former results in a better
increase of cooling rate can improve the solution quality since
performing HTLBO.
more neighborhood solutions have been exploited. At the same
According to the above analysis, all the parameters are selected
time, it is more time-consuming. In this present study, an attempt
as follows: NP = 40, T0 = 10,000, Cr = 0.85, TF = 1, Tf = 1.
has been made by considering values both 1 and 2, but no signifi-
cant difference in the results has been observed. Hence, in order to
simplify the algorithm, we take the teaching factor with the value 1 5.2. Computational results and comparisons
in our HTLBO algorithm.
In this paper, the famous Design of Experiments (DoE) approach 5.2.1. Comparisons of TLBO, VNS and HTLBO
is adopted to investigate the best parameter setting for the proposal. In this section, we compare TLBO (without local search
The value domains of these parameters are set as: NPe{20,40,60}, method), VNS and HTLBO for the PFSP with makespan criterion.
T0e{1000,5000,10,000}, Cre{0.80,0.85,0.90}, TFe{1,2}, Tf = 1, respec- For the evaluation, Carlier’s benchmark set [58] and Reeves and
tively. So, there are 2  33 = 54 combinations totally, all of which Yamada’s benchmark set [33] are used. The computational results
are tested. Teaching factor TF is the only parameter in TLBO. TF is and comparisons of TLBO (without local search method), VNS and
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 41

Fig. 4. The pseudo-code of HTLBO.

HTLBO are shown in Table 2. In Table 2, C* denotes the optimal by Zheng and Yamashiro [11]and ODDE by Li and Yin [48]. PSOVNS
solution of the instances. BPRD denotes the best percentage rela- is a algorithm which hybridize PSO and VNS. QDEA is based on the
tive difference. SD denotes the standard deviation. basic quantum-inspired evolutionary algorithm (QEA) and adopts
It is clear from Table 2 that the HTLBO algorithm is the winner, the differential evolution to perform the updating of quantum gate
since it finds the best BPRD and APRD values for all of the 29 prob- and variable neighborhood search (VNS) to raise the performance
lems, whereas the TLBO algorithm only generates 8 best BPRD val- of the local search. ODDE is a hybrid algorithm that combines DE
ues and 7 best APRD values, and the VNS algorithm yields 18 best and opposition-based search, the fast local search and pairwise
BPRD values and 17 best APRD values. The overall mean APRD val- based local search.
ues yielded by TLBO, VNS and HTLBO algorithms are equal to The computational results are shown in Table 3. As we can see
2.173%, 0.465% and 0.206%. HTLBO achieves the best performance, from Table 3, the BPRD values obtained by HTLBO are better than
which demonstrates the effectiveness of hybridization. The SD PSOVNS, QDEA and ODDE for most instances except Rec27, Rec39
obtained by HTLBO is also lower than the TLBO and VNS algo- and Rec41. The APRD values of HTLBO are also better than PSOVNS,
rithms, which justify the robustness of the HTLBO algorithm. QDEA and ODDE for most instances except Rec39. The overall
APRD values yielded by HTLBO is 0.206%, compared to the
5.2.2. Comparisons of PSOVNS, QDEA, ODDE and HTLBO corresponding values of 1.442%, 0.428% and 0.325% obtained by
The performance of HTLBO is also compared with other three PSOVNS, ODDE and QDEA, respectively. From the above observa-
state-to-art algorithms, i.e. PSOVNS by Tasgetiren et al. [40], QDEA tions, we can conclude that our HTLBO algorithm is more effective
42 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

Fig. 8. Means plot for TF factor, 95% LSD intervals.

Fig. 5. Means plot for T0 factor, 95% LSD intervals.

and efficient than PSOVNS, QDEA and ODDE algorithms. In

addition, Table 3 shows that the mean SD values resulting from
the HTLBO algorithm is smaller than PSOVNS, QDEA and ODDE
algorithms, which demonstrate the robustness of the HTLBO
algorithm.
Table 4 reports the two-side Wilcoxon rank sum tests of
HTLBO, TLBO, VNS, RSA, QDEA and ODDE algorithms with signif-
icance level equal to 5%. In the table, there are two values, i.e., p
value and h value. p is the probability of observing the given
result by chance if the null hypothesis is true. When h equals 1,
it indicates that the results obtained by the two compared algo-
rithms are obviously different. When h equals to 0, it denotes that
the difference between the two algorithms is not significant at 5%
significant level. From Table 4, we can see that the proposed
HTLBO algorithm is significantly different from the other
compared algorithms.

5.2.3. Comparisons over DMU sets for minimizing makespan

To further show the effectiveness of HTLBO, we compare HTLBO
with those of the following algorithms over the DMU benchmark
sets [5]: RSA proposed by Low et al. [59] and aforementioned
Fig. 6. Means plot for Cr factor, 95% LSD intervals.
ODDE. RSA is a robust simulated annealing algorithm. In RSA, a
mechanism that records the good solution’s characteristics is
designed and introduced into simulated annealing to make the
searching procedure more robust.
Table 5 shows the computational results and comparisons of
the HTLBO with RSA and ODDE in various testing problem size
based on DMU problems. The lower bounds and upper bounds
for these instances are also provided.
As we can see from Table 5, all of the upper bounds reported by
Demirkol et al. [5] are improved. 19 out of the 40 best known solu-
tions provided by RSA and ODDE are further improved by the pro-
posed HTLBO algorithm. Furthermore, the RSA and ODDE
algorithms found only 14 (14/40 = 35%) and 18 (18/40 = 45%) best
known solutions, respectively, while HTLBO obtained 36 (36/
40 = 90%) best known solutions. These results show that the pro-
posed HTLBO algorithm yields better solutions than all of the com-
pared algorithms.
Table 6 summarizes the average percent relative deviation
(APRD) and standard deviation (SD) obtained by RSA, ODDE and
HTLBO. The overall APRD values yielded by HTLBO is 0.315%, com-
pared to the corresponding values of 0.827% and 0.698% obtained
by RSA and ODDE, respectively. In addition, the mean SD values
Fig. 7. Means plot for NP factor, 95% LSD intervals. resulting from the HTLBO algorithm is also smaller than RSA and
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 43

Table 2
Comparisons of TLBO, VNS and HTLBO. The bold values mean the best solutions among various algorithms.

Instance n|m C* TLBO VNS HTLBO

BPRD APRD SD BPRD APRD SD BPRD APRD SD
Car1 11|5 7038 0 0 0 0 0 0 0 0 0
Car2 13|4 7166 0 0 0 0 0 0 0 0 0
Car3 12|5 7312 0 0.324 0.612 0 0 0 0 0 0
Car4 14|4 8003 0 0 0 0 0 0 0 0 0
Car5 10|6 7720 0 0.593 0.825 0 0 0 0 0 0
Car6 8|9 8505 0 0 0 0 0 0 0 0 0
Car7 7|7 6590 0 0 0 0 0 0 0 0 0
Car8 8|8 8366 0 0 0 0 0 0 0 0 0
Rec01 20|5 1247 0 0.160 0.511 0 0 0 0 0 0
Rec03 20|5 1109 0 0 0 0 0 0 0 0 0
Rec05 20|5 1242 0.242 0.242 0 0 0.242 0.125 0 0 0
Rec07 20|10 1566 0.325 0.911 1.326 0 0 0 0 0 0
Rec09 20|10 1537 0.978 1.607 1.101 0 0 0 0 0 0
Rec11 20|10 1431 1.327 1.887 1.224 0 0 0 0 0 0
Rec13 20|15 1930 0.725 1.347 1.406 0 0 0 0 0 0
Rec15 20|15 1950 0.872 2.205 1.208 0 0 0 0 0 0
Rec17 20|15 1902 2.050 3.785 1.821 0 0 0 0 0 0
Rec19 30|10 2093 2.102 3.440 1.269 0.287 0.287 0 0.143 0.215 0.182
Rec21 30|10 2017 1.636 2.677 1.483 0.645 0.694 0.603 0.149 0.199 0.143
Rec23 30|10 2011 1.542 2.984 1.926 0.249 0.448 0.420 0.099 0.174 0.089
Rec25 30|15 2513 3.104 4.497 2.036 0.438 0.637 0.536 0 0.279 0.130
Rec27 30|15 2373 2.950 4.130 2.516 0.253 0.380 0.284 0.169 0.253 0.117
Rec29 30|15 2287 4.416 6.209 1.621 0.350 0.437 0.420 0 0 0
Rec31 50|10 3045 4.696 5.813 3.142 0.657 0.936 0.760 0.263 0.263 0
Rec33 50|10 3114 2.055 3.308 1.685 0.128 0.273 0.195 0 0 0
Rec35 50|10 3277 0.335 0.458 0.821 0 0 0 0 0 0
Rec37 75|20 4951 5.777 6.461 1.842 3.252 3.979 1.201 1.272 1.495 0.873
Rec39 75|20 5087 4.030 4.993 1.581 1.710 2.025 0.982 0.944 1.238 0.751
Rec41 75|20 4960 5.161 5.907 1.621 2.722 3.448 1.546 1.290 1.855 0.981
Average 1.528 2.173 1.089 0.369 0.465 0.244 0.149 0.206 0.112

Table 3
Comparisons of PSOVNS, QDEA, ODDE and HTLBO. The bold values mean the best solutions among various algorithms.

Instance n|m C* PSOVNS QDEA ODDE HTLBO

BPRD APRD BPRD APRD BPRD APRD BPRD APRD
Car1 11|5 7038 0 0 0 0 0 0 0 0
Car2 13|4 7166 0 0 0 0 0 0 0 0
Car3 12|5 7312 0 0.420 0 0 0 0 0 0
Car4 14|4 8003 0 0 0 0 0 0 0 0
Car5 10|6 7720 0 0.039 0 0 0 0 0 0
Car6 8|9 8505 0 0.076 0 0 0 0 0 0
Car7 7|7 6590 0 0 0 0 0 0 0 0
Car8 8|8 8366 0 0 0 0 0 0 0 0
Rec01 20|5 1247 0.160 0.168 0 0 0 0 0 0
Rec03 20|5 1109 0 0.158 0 0 0 0 0 0
Rec05 20|5 1242 0.242 0.249 0.242 0.242 0 0.170 0 0
Rec07 20|10 1566 0.702 1.095 0 0 0 0 0 0
Rec09 20|10 1537 0 0.651 0 0 0 0 0 0
Rec11 20|10 1431 0.071 1.153 0 0 0 0 0 0
Rec13 20|15 1930 1.036 1.790 0.104 0.225 0 0.124 0 0
Rec15 20|15 1950 0.769 1.487 0 0.158 0 0.062 0 0
Rec17 20|15 1902 0.999 2.453 0 0.126 0 0 0 0
Rec19 30|10 2093 1.529 2.099 0.287 0.435 0.239 0.373 0.143 0.215
Rec21 30|10 2017 1.487 1.671 0.149 1.041 0.149 0.999 0.149 0.199
Rec23 30|10 2011 1.343 2.106 0.348 0.597 0.149 0.438 0.099 0.174
Rec25 30|15 2513 2.388 3.166 0.119 0.454 0.199 0.454 0 0.279
Rec27 30|15 2373 1.728 2.463 0.253 0.954 0.126 0.607 0.169 0.252
Rec29 30|15 2287 1.968 3.109 0 0.824 0 0.770 0 0
Rec31 50|10 3045 2.594 3.232 0.263 0.565 0.263 0.302 0.263 0.263
Rec33 50|10 3114 0.835 1.007 0 0.297 0 0 0 0
Rec35 50|10 3277 0 0.038 0 0 0 0 0 0
Rec37 75|20 4951 4.383 4.949 1.717 2.771 1.737 2.165 1.272 1.495
Rec39 75|20 5087 2.850 3.371 0.845 1.485 0.649 0.869 0.944 1.238
Rec41 75|20 4960 4.173 4.867 1.190 1.965 1.008 2.085 1.290 1.855
Average 1.009 1.442 0.175 0.428 0.156 0.325 0.149 0.206
44 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

Table 4
The Wilcoxon two-sided rank sum test results of HTLBO, TLBO, VNS, PSOVNS, QDEA and ODDE.

(HTLBO, TLBO) (HTLBO, VNS) (HTLBO, PSOVNS) (HTLBO, QDEA) (HTLBO, ODDE)
p h p h p h p h p h
6.21e74 1 9.32e24 1 5.21e45 1 4.78e18 1 3.06e6 1

Table 5
Best known solutions for DMU benchmark instances with makespan criterion. The bold values mean the best solutions among various algorithms.

Instance LB UB RSA ODDE HTLBO Instance LB UB RSA ODDE HTLBO

20  15 3354 4437 3899 3899 3899 40  15 5560 6986 5979 5936 5932
3168 4144 3751 3751 3751 5119 6351 5680 5670 5654
2997 3779 3518 3518 3518 5290 6506 5860 5829 5823
3420 4302 4032 4032 4032 5596 6845 5857 5830 5840
3494 4373 3910 3910 3910 5576 6783 6040 6030 6029
20  20 3776 4821 4523 4523 4523 40  15 5693 7154 6509 6486 6483
3758 4779 4424 4424 4424 5998 7528 6639 6638 6609
3902 4944 4520 4520 4520 5990 7469 6801 6768 6762
3881 4886 4496 4496 4496 6170 7608 6753 6724 6724
3823 4717 4371 4371 4371 6011 7219 6555 6530 6530
30  15 4020 5226 4543 4537 4537 50  15 6290 7673 6821 6818 6816
4080 5304 4618 4617 4609 6355 7679 6634 6651 6634
4022 5079 4547 4553 4547 6198 7416 6542 6531 6527
4490 5605 4836 4807 4807 6513 7548 6783 6787 6786
4184 5147 4757 4747 4744 6531 7750 6950 6943 6940
30  20 4806 6183 5359 5356 5346 50  20 6740 8838 7694 7673 7650
4772 6037 5654 5642 5644 6736 8539 7299 7338 7294
5004 6241 5740 5733 5715 6756 8417 7589 7581 7585
4899 6095 5440 5440 5440 6897 8590 7436 7464 7414
5757 5822 5345 5344 5342 6830 8493 7766 7753 7713

Table 6 Table 7
The computation results obtained by RSA, ODDE and HTLBO. The Wilcoxon two-sided rank sum test results of HTLBO,RSA and ODDE.
Instances RSA ODDE HTLBO (HTLBO, RSA) (HTLBO, ODDE)
APRD SD APRD SD APRD SD p h p h
20  15 0.074 0.063 0.049 0.052 0.015 0.020 4.31e32 1 2.845e25 1
20  20 0.063 0.048 0.051 0.034 0.012 0.018
30  15 1.246 0.414 0.681 0.348 0.326 0.235
30  20 1.346 0.452 0.846 0.397 0.429 0.278
40  15 1.421 0.436 0.964 0.425 0.387 0.241
40  20 1.386 0.583 0.841 0.417 0.404 0.183
50  15 0.984 0.369 1.210 0.397 0.534 0.352 improved by the proposed HTLBO algorithm. Furthermore, the
50  20 1.517 0.628 0.941 0.452 0.412 0.207 QDEA and ODDE algorithms found only and 18 (18/
Average 0.827 0.374 0.698 0.315 0.315 0.192 160 = 11.25%) and 72 (72/160 = 45%) best known solutions,
respectively, while HTLBO obtained 159 (159/160 = 99.375%) best
known solutions.
Table 9 summarizes the average percent relative deviation
ODDE algorithms. From the above observations, we can conclude (APRD) and standard deviation (SD) obtained by QDEA, ODDE
that our HTLBO algorithm outperforms RSA and ODDE algorithms and HTLBO. The overall APRD values yielded by HTLBO is 0.315%,
for solving the DMU instances with makespan criterion. compared to the corresponding values of 0.827% and 0.698%
Table 7 reports the two-side Wilcoxon rank sum tests of HTLBO, obtained by QDEA and ODDE, respectively. In addition, the mean
RSA and ODDE algorithms with significance level equal to 5%. The SD values resulting from the HTLBO algorithm is also smaller than
computational results show that the proposed HTLBO algorithm is QDEA and ODDE algorithms. These results show that the proposed
significantly different from RSA and ODDE algorithms. HTLBO algorithm outperforms all of the compared algorithms
when they are applied to solve PFSP with maximum lateness
5.2.4. Comparisons over DMU sets for minimizing maximum lateness criterion.
For the maximum lateness criterion in PFSP, we compared our Table 10 reports the two-side Wilcoxon rank sum tests of
algorithm with aforementioned ODDE and QDEA. The computa- HTLBO, QDEA and ODDE algorithms with significance level equal
tional results and comparisons of HTLBO with QDEA and ODDE to 5%. The computational results show that the proposed HTLBO
are listed in Table 8. As shown in Table 8, 88 out of the 120 best algorithm is significantly different from QDEA and ODDE
known solutions provided by QDEA and ODDE are further algorithms.
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 45

Table 8
Best solutions for DMU instances with maximum lateness criterion. The bold values mean the best solutions among various algorithms.

Instance UB QDEA ODDE HTLBO Problem UB QDEA ODDE HTLBO

20  15 20  20
1 2833 2468 2431 2431 21 3437 3024 2957 2957
2 2322 2087 2059 2059 22 3127 2752 2716 2716
3 2370 2112 2065 2065 23 2906 2745 2737 2737
4 2554 2275 2230 2230 24 3197 2995 2927 2927
5 2699 2330 2321 2321 25 3069 2748 2706 2706
6 2239 2307 2307 2307 26 2594 2579 2579 2579
7 1722 1712 1712 1712 27 3388 3294 3294 3294
8 2526 2508 2508 2508 28 2978 2947 2947 2947
9 2165 2132 2132 2132 29 2271 2210 2210 2210
10 2292 2345 2345 2340 30 2836 2740 2740 2740
11 3360 3042 3012 3012 31 3878 3652 3616 3616
12 3651 3212 3182 3182 32 3914 3564 3526 3526
13 3318 2908 2881 2881 33 4076 3683 3638 3638
14 3347 3092 3037 3037 34 4276 3931 3865 3859
15 3251 3049 3012 3012 35 3853 3580 3545 3545
16 3009 2589 2563 2563 36 3231 3115 3080 3080
17 2892 2627 2600 2600 37 3279 3095 3053 3053
18 2462 2330 2302 2302 38 3514 3364 3347 3347
19 2635 2531 2518 2518 39 2998 2975 2975 2975
20 2533 2457 2436 2436 40 3370 3188 3188 3188
30  15 30  20
41 2837 2291 2206 2206 61 3737 3207 3107 3088
42 3088 2629 2506 2493 62 3592 3065 3008 2981
43 2733 2346 2249 2238 63 4115 3496 3361 3332
44 3054 2689 2513 2513 64 3731 3414 3292 3292
45 3074 2636 2587 2492 65 3254 2894 2823 2803
46 2158 2026 2007 2007 66 3296 3191 3191 3191
47 1875 1748 1740 1740 67 3057 2934 2934 2934
48 2637 2591 2575 2575 68 3158 3137 3137 3137
49 2366 2333 2288 2288 69 3134 3166 3166 3166
50 2381 2368 2360 2360 70 1994 1941 1893 1893
51 4465 3750 3657 3646 71 4472 4007 3944 3929
52 4197 3583 3504 3501 72 4603 4199 4084 4081
53 3810 3236 3171 3166 73 4884 4430 4328 4314
54 4472 3865 3791 3787 74 4628 4332 4199 4161
55 4270 3631 3550 3513 75 4678 4117 4052 4046
56 3221 2792 2671 2669 76 3997 3708 3630 3612
57 2983 2679 2596 2593 77 3721 3461 3366 3366
58 3279 2876 2804 2790 78 3591 3370 3312 3312
59 3433 3053 2982 2971 79 4178 3877 3831 3806
60 3252 2976 2896 2896 80 4111 3936 3853 3853
40  15 40  20
81 3530 2619 2430 2430 101 4336 3495 3368 3334
82 3355 2662 2522 2496 102 4278 3713 3593 3539
83 3312 2728 2623 2605 103 4216 3610 3470 3458
84 3060 2552 2430 2407 104 4139 3682 3402 3463
85 3159 2691 2598 2555 105 4078 3612 3474 3436
86 2584 2370 2343 2337 106 3379 3132 3132 3132
87 2343 2112 2088 2064 107 3236 3212 3085 3085
88 2364 2364 2364 2364 108 2891 2801 2779 2779
89 2364 2375 2375 2375 109 3627 3339 3303 3303
90 2503 2419 2419 2419 110 2610 2505 2505 2505
91 5152 4426 4328 4236 111 5438 4842 4730 4624
92 4859 3932 3850 3800 112 5640 4943 4775 4754
93 4969 4441 4280 4270 113 5873 5066 4964 4935
94 4854 4123 4096 3951 114 5560 4977 4855 4843
95 5133 4391 4274 4270 115 5536 4954 4854 4758
96 3596 3198 3152 3089 116 4177 3643 3534 3500
97 3470 3245 3097 3067 117 4066 3707 3571 3541
98 3464 3145 2970 2935 118 4590 4030 3852 3831
99 3479 3209 3103 3073 119 3953 3757 3524 3474
100 3299 3021 2863 2848 120 4320 3946 3796 3765
50  15 50  20
121 4016 2976 2809 2777 141 4495 3659 3533 3426
122 3821 2935 2739 2710 142 4713 3755 3589 3574
123 3745 2774 2638 2593 143 4262 3544 3384 3361
124 3631 2774 2621 2612 144 4922 3983 3845 3735
125 3769 2942 2859 2722 145 4380 3608 3424 3402
126 2771 2602 2602 2602 146 3654 3629 3513 3482
127 2979 2972 2969 2969 147 2816 2779 2734 2734
128 3276 3064 3064 3064 148 3593 3502 3464 3464
129 2615 2606 2581 2581 149 3812 3632 3605 3604

(continued on next page)

46 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47

Table 8 (continued)

Instance UB QDEA ODDE HTLBO Problem UB QDEA ODDE HTLBO

130 3211 3190 3190 3190 150 3596 3672 3506 3506
131 5364 4554 4332 4324 151 6224 5462 5270 5253
132 5944 4765 4662 4603 152 6582 5749 5547 5522
133 5294 4540 4446 4406 153 6462 5752 5572 5518
134 5538 4659 4518 4427 154 6074 5576 5465 5401
135 5226 4516 4341 4300 155 6166 5285 5174 5154
136 3817 3433 3283 3230 156 4472 4098 3914 3858
137 3866 3473 3335 3217 157 4438 4123 3985 3939
138 3843 3483 3313 3251 158 4461 4193 4013 3910
139 4007 3233 3089 3065 159 4259 3958 3815 3745
140 3997 3806 3610 3519 160 4521 4110 3971 3933

Table 9
The computation results obtained by QDEA, ODDE and HTLBO. Acknowledgments
Instances QDEA ODDE HTLBO
This research is supported by the State Key Program of National
APRD SD APRD SD APRD SD Natural Science of China (Grant No. 51035001), National Science
20  15 2.034 0.421 0.384 0.128 0.108 0.067 Foundation of China (Grant No. 51275190) and National High
20  20 2.318 0.517 0.417 0.249 0.117 0.052 Technology Research and Development Program of 863 projects
30  15 4.671 0.814 1.842 0.608 0.398 0.241
(Grant No. 2012AA040909).
30  20 2.975 0.632 1.126 0.521 0.542 0.312
40  15 5.321 1.024 1.536 0.537 0.483 0.348
40  20 5.418 1.226 0.973 0.412 0.528 0.287
50  15 6.024 1.171 1.265 0.643 0.367 0.231
50  20 5.699 0.789 1.452 0.428 0.425 0.182 References
Average 4.308 0.824 1.124 0.441 0.371 0.215
[1] Framinan JM, Gupta JN, Leisten R. A review and classification of heuristics for
permutation flow-shop scheduling with makespan objective. J Oper Res Soc
2004;55(12):1243–55.
[2] Adenso-Díaz B. An SA/TS mixture algorithm for the scheduling tardiness
problem. Eur J Oper Res 1996;88(3):516–24.
Table 10 [3] Framinan JM, Leisten R. Total tardiness minimization in permutation flow
The Wilcoxon two-sided rank sum test results of HTLBO,QDEA and ODDE. shops: a simple approach based on a variable greedy algorithm. Int J Prod Res
2008;46(22):6479–98.
(HTLBO, QDEA) (HTLBO, ODDE) [4] Vallada E, Ruiz R. Genetic algorithms with path relinking for the minimum
p h p h tardiness permutation flowshop problem. Omega 2010;38(1):57–67.
[5] Demirkol E, Mehta S, Uzsoy R. Benchmarks for shop scheduling problems. Eur J
3.042e51 1 1.728e32 1 Oper Res 1998;109(1):137–41.
[6] Blazewicz J, Lenstra JK, Kan A. Scheduling subject to resource constraints:
classification and complexity. Discrete Appl Math 1983;5(1):11–24.
[7] Lenstra JK, Kan AR, Brucker P. Complexity of machine scheduling problems.
Ann Discrete Math 1977;1:343–62.
6. Conclusion and future research [8] Bansal S. Minimizing the sum of completion times of n jobs over m machines
in a flowshop—a branch and bound approach. AIIE Trans 1977;9(3):306–11.
This paper proposes a novel hybrid Teaching–Learning-Based [9] Della Croce F, Narayan V, Tadei R. The two-machine total completion time flow
shop problem. Eur J Oper Res 1996;90(2):227–37.
Optimization (HTLBO) algorithm to address the permutation flow [10] Ignall E, Schrage L. Application of the branch and bound technique to some
shop scheduling problem (PFSP) with the makespan and maximum flow-shop scheduling problems. Oper Res 1965;13(3):400–12.
lateness criterion. In order to employ general TLBO to solve PFSP [11] Zheng T, Yamashiro M. Solving flow shop scheduling problems by quantum
differential evolutionary algorithm. Int J Adv Manuf Technol 2010;49(5–
successfully, a LOV rule is used to convert the individual to the
8):643–62.
job permutation. VNS is combined with TLBO to achieve the bal- [12] Rao R, Savsani V, Vakharia D. Teaching–learning-based optimization: a novel
ance of diversification and intensification. Furthermore, five kinds method for constrained mechanical design optimization problems. Comput
of neighborhood structures are introduced in VNS to obtain prom- Aided Des 2011;43(3):303–15.
[13] Rao R, Savsani V, Vakharia D. Teaching–learning-based optimization: an
ising results, and a re-initialization mechanism is designed to optimization method for continuous non-linear large scale problems. Inf Sci
increase the probability of escaping from the local optimum. 2012;183(1):1–15.
The computational results and comparisons based on Carlier’s [14] Gupta JN, Stafford Jr EF. Flowshop scheduling research after five decades. Eur J
Oper Res 2006;169(3):699–711.
benchmark set and Reeves and Yamada’s benchmark set show [15] Tseng FT, Stafford EF, Gupta JN. An empirical analysis of integer programming
the effectiveness of the proposed algorithm. HTLBO is also applied formulations for the permutation flowshop. Omega 2004;32(4):285–93.
to solve the well-known DMU problems, and compare with the [16] Palmer D. Sequencing jobs through a multi-stage process in the minimum
total time – a quick method of obtaining a near optimum. OR 1965:101–7.
algorithms of the recent literature. For makespan criterion, the [17] Campbell HG, Dudek RA, Smith ML. A heuristic algorithm for the n job, m
HTLBO algorithm has been tested against the other 5 well perform- machine sequencing problem. Manage Sci 1970;16(10):B630–7.
ing algorithms from the recent literature. For maximum lateness [18] Dannenbring DG. An evaluation of flow shop sequencing heuristics. Manage
Sci 1977;23(11):1174–82.
criterion, the HTLBO algorithm has been tested against 2 recent [19] Nawaz M, Enscore EE, Ham I. A heuristic algorithm for the m-machine, n-job
algorithms. The experimental results show that the proposed flow-shop sequencing problem. Omega 1983;11(1):91–5.
HTLBO method performs well and is an effective approach for [20] Taillard E. Some efficient heuristic methods for the flow shop sequencing
problem. Eur J Oper Res 1990;47(1):65–74.
the PFSP with the makespan and maximum lateness criterion.
[21] Framinan J, Leisten R. An efficient constructive heuristic for flowtime
Our algorithm improves the best known solutions for 19 instances minimisation in permutation flow shops. Omega 2003;31(4):311–7.
with makespan criterion and 88 instances with maximum lateness [22] Framinan JM, Leisten R, Ruiz-Usano R. Efficient heuristics for flowshop
criterion from the well-known DMU benchmark. The future sequencing with the objectives of makespan and flowtime minimisation. Eur
J Oper Res 2002;141(3):559–69.
research directions involve the consideration of multi-objective [23] Li X, Wang Q, Wu C. Efficient composite heuristics for total flowtime
flow shop and hybrid flow shop scheduling problems. minimization in permutation flow shops. Omega 2009;37(1):155–64.
 Z. Xie et al. / Advances in Engineering Software 77 (2014) 35–47 47

[24] Laha D, Chakraborty UK. An efficient hybrid heuristic for makespan [41] Pan Q, Tasgetiren MF, Liang Y. A discrete differential evolution algorithm for
minimization in permutation flow shop scheduling. Int J Adv Manuf Technol the permutation flowshop scheduling problem. Comput Ind Eng
2009;44(5–6):559–69. 2008;55(4):795–816.
[25] Ruiz R, Maroto C. A comprehensive review and evaluation of permutation [42] Chang P, Huang W, Ting C, Wu L, Lai C. A hybrid genetic-immune algorithm
flowshop heuristics. Eur J Oper Res 2005;165(2):479–94. with improved offsprings and elitist antigen for flow-shop scheduling
[26] Osman I, Potts C. Simulated annealing for permutation flow-shop scheduling. problems. High Performance Computing and Communications, 2009
Omega 1989;17(6):551–7. HPCC’09. In: 11th IEEE International conference on: IEEE; 2009. p. 591–6.
[27] Ogbu F, Smith DK. The application of the simulated annealing algorithm to the [43] Murata T, Ishibuchi H, Tanaka H. Genetic algorithms for flowshop scheduling
solution of the n/m/Cmax flowshop problem. Comput Oper Res problems. Comput Ind Eng 1996;30(4):1061–71.
1990;17(3):243–53. [44] Nearchou AC. A novel metaheuristic approach for the flow shop scheduling
[28] Reeves CR. Improving the efficiency of tabu search for machine sequencing problem. Eng Appl Artif Intell 2004;17(3):289–300.
problems. J Oper Res Soc 1993:375–82. [45] Ahmadizar F. A new ant colony algorithm for makespan minimization in
[29] Nowicki E, Smutnicki C. A fast tabu search algorithm for the permutation flow- permutation flow shops. Comput Ind Eng 2012;63(2):355–61.
shop problem. Eur J Oper Res 1996;91(1):160–75. [46] Tzeng Y-R, Chen C-L, Chen C-L. A hybrid EDA with ACS for solving permutation
[30] Watson J, Barbulescu L, Whitley LD, Howe AE. Contrasting structured and flow shop scheduling. Int J Adv Manuf Technol 2012;60(9–12):1139–47.
random permutation flow-shop scheduling problems: search-space topology [47] Allahverdi A. Two-machine proportionate flowshop scheduling with breakdowns
and algorithm performance. INFORMS J Comput 2002;14(2):98–123. to minimize maximum lateness. Comput Oper Res 1996;23(10):909–16.
[31] Grabowski J, Wodecki M. A very fast tabu search algorithm for the [48] Li X, Yin M. An opposition-based differential evolution algorithm for
permutation flow shop problem with makespan criterion. Comput Oper Res permutation flow shop scheduling based on diversity measure. Adv Eng
2004;31(11):1891–909. Softw 2013;55:10–31.
[32] Reeves CR. A genetic algorithm for flowshop sequencing. Comput Oper Res [49] Satapathy SC, Naik A. Data clustering based on teaching–learning-based
1995;22(1):5–13. optimization. Swarm, evolutionary, and memetic computing. Springer; 2011.
[33] Reeves CR, Yamada T. Genetic algorithms, path relinking, and the flowshop p. 148–56.
sequencing problem. Evol Comput 1998;6(1):45–60. [50] Toğan V. Design of planar steel frames using teaching–learning based
[34] Rajendran C, Ziegler H. Ant-colony algorithms for permutation flowshop optimization. Eng Struct 2012;34:225–32.
scheduling to minimize makespan/total flowtime of jobs. Eur J Oper Res [51] Venkata Rao R, Patel V. Multi-objective optimization of two stage
2004;155(2):426–38. thermoelectric cooler using a modified teaching–learning-based
[35] Dong X, Huang H, Chen P. An iterated local search algorithm for the optimization algorithm. Eng Appl Artif Intell 2013;26(1):430–45.
permutation flowshop problem with total flowtime criterion. Comput Oper [52] Keesari H, Rao R. Optimization of job shop scheduling problems using
Res 2009;36(5):1664–9. teaching–learning-based optimization algorithm. OPSEARCH 2013:1–17.
[36] Ruiz R, Stützle T. A simple and effective iterated greedy algorithm for the [53] Bean JC. Genetic algorithms and random keys for sequencing and
permutation flowshop scheduling problem. Eur J Oper Res optimization. ORSA J Comput 1994;6(2):154–60.
2007;177(3):2033–49. [54] Qian B, Wang L, Hu R, Wang W-L, Huang D-X, Wang X. A hybrid differential
[37] Pan Q-K, Ruiz R. Local search methods for the flowshop scheduling problem evolution method for permutation flow-shop scheduling. Int J Adv Manuf
with flowtime minimization. Eur J Oper Res 2012;222(1):31–43. Technol 2008;38(7–8):757–77.
[38] Tasgetiren MF, Liang Y, Sevkli M, Gencyilmaz G. Particle swarm optimization [55] Glodberg DE. Genetic algorithms in search, optimization, and machine
algorithm for makespan and maximum lateness minimization in permutation learning. Addison Wesley; 1989.
flowshop sequencing problem. In: Proceedings of the fourth international [56] Hansen P, Mladenović N. Variable neighborhood search: principles and
symposium on intelligent manufacturing systems, Sakarya, Turkey; 2004. p. applications. Eur J Oper Res 2001;130(3):449–67.
431–41. [57] Kirkpatrick Jr S, Gelatt CD, Vecchi MP. Optimization by simulated annealing.
[39] Liao C, Tseng C, Luarn P. A discrete version of particle swarm optimization for Science 1983;220(4598):671–80.
flowshop scheduling problems. Comput Oper Res 2007;34(10):3099–111. [58] Carlier J. Ordonnancements a contraintes disjonctives. RAIRO-Operations
[40] Tasgetiren MF, Liang Y, Sevkli M, Gencyilmaz G. A particle swarm optimization Research-Recherche Opérationnelle. 1978;12(4):333–50.
algorithm for makespan and total flowtime minimization in the permutation [59] Low C, Yeh J, Huang K. A robust simulated annealing heuristic for flow shop
flowshop sequencing problem. Eur J Oper Res 2007;177(3):1930–47. scheduling problems. Int J Adv Manuf Technol 2004;23(9–10):762–7.